end-to-end learning for combining multiple robot actions · end-to-end learning for combining...

18
End-to-end learning for combining multiple robot actions February 14 th , 2018 Tetsuya Ogata Joint Appointed Fellow, AIST AI Research Center Professor, Waseda University

Upload: others

Post on 13-Mar-2020

9 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: End-to-end learning for combining multiple robot actions · End-to-end learning for combining multiple robot actions February 14th, 2018 Tetsuya Ogata Joint Appointed Fellow, AIST

������ �������� ����������

End-to-end learning for combining multiple robot

actionsFebruary 14th, 2018

Tetsuya Ogata

Joint Appointed Fellow, AIST AI Research Center

Professor, Waseda University

Page 2: End-to-end learning for combining multiple robot actions · End-to-end learning for combining multiple robot actions February 14th, 2018 Tetsuya Ogata Joint Appointed Fellow, AIST

A H

������•

���••

Page 3: End-to-end learning for combining multiple robot actions · End-to-end learning for combining multiple robot actions February 14th, 2018 Tetsuya Ogata Joint Appointed Fellow, AIST

c S a c[ S ] cO

D E CE E C G C. A F C G

KGE G , GHE F

1C A F C G

) FF E

2 1 E

3HGDHG ) FF F

. A F C G

D 1 E

3HGDHG ) FF F

A E C G C ( 0E M IF L GD E C G C - G

Page 4: End-to-end learning for combining multiple robot actions · End-to-end learning for combining multiple robot actions February 14th, 2018 Tetsuya Ogata Joint Appointed Fellow, AIST

I PMO SL FH E

/ 4 / //

( D +

/ C /

/ /

/ 4

+ /

/

+ D

( D+ )

/

+ D

Page 5: End-to-end learning for combining multiple robot actions · End-to-end learning for combining multiple robot actions February 14th, 2018 Tetsuya Ogata Joint Appointed Fellow, AIST

���� �� �������� �����������

Cognitive Robotics- High caliculation

cost

- Redesign for each task

ComplexTask

Applicability

< Motion Generation Method for Robotics >Robotics with Deep Learning

[K.Yamazaki,2012]

[K.Noda, 2014]

- Low caliculation cost

- Same architecture

- Industrial robot

- Hight-DOF robot

[S.Levine, 2016]

Our Studies

[P.C.Yang, 2016]

Design model Guided Policy Learning

- Extensive search time

- Redesign for each task

- Experimentalrobot

- Low-DOF

Page 6: End-to-end learning for combining multiple robot actions · End-to-end learning for combining multiple robot actions February 14th, 2018 Tetsuya Ogata Joint Appointed Fellow, AIST

O PK RL YTM S K I

2A62

6

2 2A62

6

6 A C6 62 6 A 2A

12 2 2 E 2 6 2 2 0 , 2A2

. A 2A 2A 6AA6

Page 7: End-to-end learning for combining multiple robot actions · End-to-end learning for combining multiple robot actions February 14th, 2018 Tetsuya Ogata Joint Appointed Fellow, AIST

G M OI J SR DF C

<

,> >>

Teleoperation DCAE extract feature vector from raw image

>

, >+

<< >

- 3D mouse and HMD

- Applicable to non-backdrivable robot

-

< <

MTRNNlearn sequencesby hierarchizing

B< < > >

Page 8: End-to-end learning for combining multiple robot actions · End-to-end learning for combining multiple robot actions February 14th, 2018 Tetsuya Ogata Joint Appointed Fellow, AIST

gh efa gh cbdghT S

• 3 : -) 7PG F CF F 7

• 2F F -1 0FC F FD GC : :CF5 F FD GF R

0F G G F DD

• 7 -5 : D D - && O&& O(

(,+( DD GG F C &) D

6 DGC & 046 ( : M , :

7 7

CF 0FC 7

Page 9: End-to-end learning for combining multiple robot actions · End-to-end learning for combining multiple robot actions February 14th, 2018 Tetsuya Ogata Joint Appointed Fellow, AIST

C I F

- -- -

DCAE Encoding DCAE Decoding

TDNN -

-

-

MTRNN

Page 10: End-to-end learning for combining multiple robot actions · End-to-end learning for combining multiple robot actions February 14th, 2018 Tetsuya Ogata Joint Appointed Fellow, AIST

L TOH P

6 60 6 1060: 6 ??? = =.0 / ? / :

Page 11: End-to-end learning for combining multiple robot actions · End-to-end learning for combining multiple robot actions February 14th, 2018 Tetsuya Ogata Joint Appointed Fellow, AIST

@ C 7

1

1 20

Page 12: End-to-end learning for combining multiple robot actions · End-to-end learning for combining multiple robot actions February 14th, 2018 Tetsuya Ogata Joint Appointed Fellow, AIST

C E R ONP A

2:

, 2 1 2 2 :2

2 1 , 2 2 1:2 1 ,

:

2 2 2 : 2

2

K. Kase, K. Suzuki, P. Yang, H. Mori, and T. Ogata: IEEE ICRA2018 (accepted)

Page 13: End-to-end learning for combining multiple robot actions · End-to-end learning for combining multiple robot actions February 14th, 2018 Tetsuya Ogata Joint Appointed Fellow, AIST

MC M I M :

2 1

1 1 12 2

1 1

1 2 1

Page 14: End-to-end learning for combining multiple robot actions · End-to-end learning for combining multiple robot actions February 14th, 2018 Tetsuya Ogata Joint Appointed Fellow, AIST

R H R R A

2 2 2 2 22 2 2 2 2 2

2 2

: :

Page 15: End-to-end learning for combining multiple robot actions · End-to-end learning for combining multiple robot actions February 14th, 2018 Tetsuya Ogata Joint Appointed Fellow, AIST

R

5 81 1221/ 5:0 5-8 .

4 .8 3 : 505- / 9 .8 3. 5/ 3 / - -:

4 . - 5:2 01:99-5 - 4 98

4 9 : 5 - 9- 75 / 9: - 5/81 :1 4 98

4 :1 9 :- 5 - 5/81-

Page 16: End-to-end learning for combining multiple robot actions · End-to-end learning for combining multiple robot actions February 14th, 2018 Tetsuya Ogata Joint Appointed Fellow, AIST

NS R O VP X W OZ KM

A DEC E A D AF E A A AE

F E E D D I E EC E FDD D C BCA C

E A EAH ,FEE E D EA AH

Page 17: End-to-end learning for combining multiple robot actions · End-to-end learning for combining multiple robot actions February 14th, 2018 Tetsuya Ogata Joint Appointed Fellow, AIST

OT VSUP W R X P Z NMK

D F FGD C G F AC C CF

1I B 1 F CB B F

B CIE

00 000

CBFDC DC EE C B F H DC EE

0 0

F D 1

Page 18: End-to-end learning for combining multiple robot actions · End-to-end learning for combining multiple robot actions February 14th, 2018 Tetsuya Ogata Joint Appointed Fellow, AIST

PR TN V U I

? A

F . ? ?? ? ?

?. H

F ? ? A ?A A ? ? ?

? ? ))F ? ? ? A ? ( ,

- ? ?